While most vulnerability detection methods rely on source code features, such as Abstract Syntax Tree (AST), some have adopted compiler intermediate representations (IRs) for their efficient code representations. However, IR-based methods struggle with modeling long-range dependencies, exhibit limitations in semantic extraction, and suffer from inadequate model evaluation. To address these challenges, we propose WelkIR, a two-tier IR-based Transformer model for vulnerability detection, which is fine-tuned from a flow-sensitive pre-trained model on task-specific data. During pre-training, we introduce three novel pre-training tasks to capture the program semantics: Control Flow Prediction (CFP), Data Def-Use Prediction (DDP), and Data Reachability Prediction (DRP). The pre-trained model is then fine-tuned on task-specific data to refine its understanding of vulnerable code patterns, resulting in a specialized vulnerability detection model. We evaluate WelkIR through comprehensive experiments. On the Juliet benchmark, WelkIR achieves a 7.01%–23.26% improvement in macro F1-score compared to state-of-the-art baselines. We additionally construct ARVul, which, to the best of our knowledge, is the first large-scale LLVM IR-based vulnerability dataset, comprising 3,732 vulnerable functions from 217 real-world projects. On ARVul, WelkIR outperforms baselines by 9.01%–32.74% in F1-score. Ablation studies confirm that the proposed pre-training tasks contribute to WelkIR’s performance in vulnerability detection.